Student Internship Admissions

Introduction

Demographic Insights

Geographic

Code

import pandas as pd
import altair as alt
import seaborn as sns
import plotly.express as px
from vega_datasets import data
import matplotlib.pyplot as plt

Code

df = pd.read_csv('../data/clean_data.csv')

Code

custom_palette = ['#00072D', '#0A2472', '#0E6BA8', '#A6E1FA', '#99ABC5', '#8B748F', '#6F0624']

Code

#calculate averages of all numeric columns
num_cols = df[['State', 'GPA', 'WorkExp', 'TestScore', 'WritingScore', 'VolunteerLevel']]
avg_df = num_cols.groupby('State').mean().reset_index()
state_abbr = {
    'Alabama': 'AL',
    'California': 'CA',
    'Colorado': 'CO',
    'Florida': 'FL',
    'Georgia': 'GA',
    'Mississippi': 'MS',
    'New York': 'NY',
    'Oregon': 'OR',
    'Utah': 'UT',
    'Vermont': 'VT',
    'Virginia': 'VA'

}
avg_df['State_Abbr'] = avg_df['State'].map(state_abbr)
avg_df = avg_df.drop(columns=['State'])
avg_df = avg_df.rename(columns={'State_Abbr': 'State'})

Code

import plotly.io as pio
pio.renderers.default = "plotly_mimetype+notebook"

fig = px.choropleth(avg_df, locationmode="USA-states", 
                    locations=avg_df['State'], 
                    scope="usa",
                    color=avg_df['GPA'],
                    hover_data={"State": True, "GPA": True},  
                    labels={"GPA": "Selected Variable"},
                    color_continuous_scale=custom_palette
                )

dropdown = []
for col in avg_df.columns[:-1]:
    dropdown.append({'label': col, 'method': 'update', 'args': [{'z': [avg_df[col]]}]})

fig.update_layout(updatemenus=[{'buttons': dropdown, 'direction': 'down', 'showactive': True}],
                  title='Choropleth Map of Average Selected Variable')
fig.update_coloraxes(colorbar_title=dict(text='Selected Variable'))

fig.show()

Code

#create dataframe of rates for each state by decision
decision_state = df.groupby(['Decision', 'State'])[["GPA"]].count().reset_index()
decision_state = decision_state.rename(columns={'GPA':'StateCount'})
decision_state['DecisionCount'] = decision_state.groupby('Decision')['StateCount'].transform('sum')
decision_state['Rate'] = decision_state['StateCount'] / decision_state['DecisionCount'] * 100

state_id_dict = dict(zip(data.population_engineers_hurricanes()["state"], data.population_engineers_hurricanes()["id"]))
decision_state["StateID"] = decision_state["State"].map(state_id_dict)

admit_states = decision_state[decision_state['Decision'] == "Admit"]
decline_states = decision_state[decision_state['Decision'] == "Decline"]

Code

states = alt.topo_feature('https://raw.githubusercontent.com/vega/vega-datasets/master/data/us-10m.json', 'states')
click = alt.selection_multi(fields = ["State"])

existing_states = alt.Chart(states).mark_geoshape(stroke='black').encode(
    color = alt.Color("Rate:Q", scale=alt.Scale(range=custom_palette)),
    tooltip = ["State:N", "Rate:Q"],
    opacity = alt.condition('isValid(datum.Rate)', alt.value(1), alt.value(0.2)),
).transform_lookup(
    lookup = "id",
    from_ = alt.LookupData(admit_states, "StateID", list(admit_states.columns))
).properties(width = 333, height = 200, title="Admission Rates by State").add_selection(click).project(type = "albersUsa").interactive()

missing_states = (
    alt.Chart(states)
    .mark_geoshape(fill = "grey", stroke = "white")
    .encode(opacity=alt.condition("isValid(datum.Rate)", alt.value(0), alt.value(0.2))).add_selection(click).project(type = "albersUsa")
    )

admit_map = existing_states + missing_states
admit_map = admit_map.encode(
    tooltip= ["State:N", "Rate:Q"]
    ).transform_lookup(
        lookup="id",
        from_=alt.LookupData(admit_states, "StateID", list(admit_states.columns))
    ).interactive()

existing_states = alt.Chart(states).mark_geoshape(stroke='black').encode(
    color = alt.Color("Rate:Q", scale=alt.Scale(range=custom_palette)),
    tooltip = ["State:N", "Rate:Q"],
    opacity = alt.condition('isValid(datum.Rate)', alt.value(1), alt.value(0.2)),
).transform_lookup(
    lookup = "id",
    from_ = alt.LookupData(decline_states, "StateID", list(decline_states.columns))
).properties(width = 333, height = 200, title="Rejection Rates by State").add_selection(click).project(type = "albersUsa").interactive()

missing_states = (
    alt.Chart(states)
    .mark_geoshape(fill = "grey", stroke = "white")
    .encode(opacity=alt.condition("isValid(datum.Rate)", alt.value(0), alt.value(0.2))).add_selection(click).project(type = "albersUsa")
    )

decline_map = existing_states + missing_states
decline_map = decline_map.encode(
    tooltip= ["State:N", "Rate:Q"]
    ).transform_lookup(
        lookup="id",
        from_=alt.LookupData(decline_states, "StateID", list(decline_states.columns))
    ).interactive()

admit_map | decline_map

Gender

Machine Learning

Before applying machine learning, below we will explain the reasoning behind applying ML to the student admissions data. We wanted to identify relationships between the academic features in the data on the decision, namely GPA, writing score, and test score. The motivation behind this is to use the information of the relationships, if any are present, to help students in our university understand what features of their application might contribute to the decision. We then hope to provide targeted help to our students based on our findings to increase the number of students admitted to internships.

Code

%%html
<img src="../website/images/decision_pairplot.png">

Above is the pairplot of GPA, writing score, and test score of the students grouped by the decision. When looking at the scatterplots, we notice some patterns:

• Students with low test score, no matter the GPA, were declined.
• Students with high test score and high GPA were accepted.
• Students with a pretty high GPA but average test score were waitlisted.
• Students with high test score, no matter the writing score, were admitted.
• Students with a low test score, no matter the writing score, were declined.
• Students with high writing scores but average test score were waitlisted.

Through the pairplot we can see that some of the academic features have relationships by decision result, but some features seem to be more important than others. We will investigate the importance of all features on decision using XGBoost and Shapley values.

XGBoost and Shapley

Conclusions